Early warning systems (EWS) that predict patient outcomes for therapeutic interventions are increasingly being used to make clinical decisions. Predicted events often include outcomes such as unplanned ICU transfer, cardiac arrest, or death. The goal of these models is to provide a timely alert to clinicians so they can intervene and prevent these outcomes from occurring. However, it is difficult to evaluate the performance of an EWS after it has been deployed. For example, if the model correctly predicts a deterioration event, the clinician may take actions to prevent the patient’s predicted deterioration. If the clinician’s actions are successful, the deterioration is averted and the label against which the model is tested will become negative. This is despite the fact that the model’s positive prediction was correct. Without the context of the clinician intervention, the model will be evaluated as if it was wrong. Overall, the better the model is at improving outcomes, the worse the performance is going to appear. This is known as confounding medical interventions (CMI). In this project, we evaluated two causal inference frameworks for estimating model performance metrics such as sensitivity and PPV in the presence of CMI. To our knowledge, no other methods exist for estimating the performance of a model in the context of confounding medical interventions.
I started doing research at the Medical College of Wisconsin while getting my Master's in Bioinformatics focusing on identifying genomic loci associated with diabetic traits in rats. I was really excited to be a part of a great group of people figuring out an interesting and complex problem. I continued doing genomics research throughout my Ph.D. at Marquette University and post-doc at the University of Wisconsin-Madison. The majority of my work as a post-doc focused on understanding genetic variation in the context of the 3D genome. The spatial organization of the genome is such an important component of gene regulation, but at the time was previously understudied. I was involved in a number of collaborations involving this, and it was awesome being able to work with such crucial datasets (and it was A LOT of data). Coming to Michigan Medicine and working more on clinical problems rather than genomics has been a bit of a shift, but I have been able to apply all the research skills I gained over the last 10 years and continue to be excited by new and interesting data science problems.
I think the most exciting thing about data science and AI research is not only using models for prediction but also what features drive the prediction. I also appreciate the increased focus on addressing potential bias in AI and making models more equitable and fairer. I am a big fan of looking "underneath the hood" and I am excited that more people are going beyond simply looking at overall performance. The picture is often much more complex and there are many nuances and potential pitfalls. However, addressing them is part of the process and can lead to increased public trust and model accuracy and usage.
I run 800-1,200 miles total every year, but not I'm not at all fast.
I have two Shiba Inus named after food (Marzipan and Soba).